The role of AI in product development

Ryan J. Salva, VP of Product at GitHub, discusses how GitHub Copilot went from an experimental idea to a generally available AI pair-programming tool. The conversation covers the origins of the project with OpenAI, how the GitHub Next research team handed off to a product team, handling ethical and scaling challenges, and his thinking on the future of AI in software development.

01Key learnings

• Ring-fence a dedicated R&D team for long-horizon, uncertain bets, reserving roughly 5-10% of capacity for audacious experiments alongside operations and incremental product work.
• Hand off research prototypes only after customer signal exists, and move researchers back to R&D based on a trained replacement in seat, not a calendar date.
• Let the team that owns the product day to day control its own roadmap; don't outsource innovation entirely to an R&D group.
• Introduce engineering fundamentals like security, privacy, uptime and accessibility gradually as an experiment graduates into an operational product, and expect cultural change.
• Tune latency to the flow-state threshold; Copilot's team found roughly 200 milliseconds as the sweet spot before suggestions feel interruptive.
• Frame AI as a pair programmer with clear behavioral principles, and involve legal, privacy, security and users when deciding content filtering, since block lists are fraught.
• Spend heavy time with customers and the community, especially on contested issues like training data, security and accountability, and scale up product staff accordingly.
• Position AI as an augmenter that keeps existing tests, static analysis and human review in place rather than a replacement for developers.

02Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.